Crash when call ov::core.comile_model

[Jsy0220]

Hi
I upgraded OpenVino from 2022.1 to 2024.0 for my application recently. Following are the situation:

1. Always crash at ov::core.compile_model no matter what model loaded when test in unittest.
2. both official lib and locally compiled lib have been tried, same crash.
3. when the application compiled as a dynamic lib loaded in another application, there was no crash.

Any ideas ?
Thank you so much !!

[ilya-lavrenov]

Hi @Jsy0220
We need a reproducer, as we cannot magically guess what happened.

[Jsy0220]

@ilya-lavrenov hi, I found that when the application links both OpenVino and tbb libs, it would happen. Specifically these two:

I guess there may be a symbol conflict ？

[ilya-lavrenov]

it's hard to guess without code reproducer.

[Jsy0220]

@ilya-lavrenov Sorry, there is no special code to reproduce it, just link these two libs and load a model to run ov::core.compile_model() like example hello_classification

[rang-inspekto]

I get a similar issue when using libtcmalloc.so.4.
Using OpenVino 2024.6, Tensorflow 2.18, Python 3.11 (installing from pip).

simple code to reproduce:

import numpy as np
import openvino as ov
from openvino.runtime import Core
import tensorflow as tf
from tensorflow.keras.applications import ResNet101
from tensorflow.keras.layers import Input

tf.config.experimental.set_visible_devices([], 'GPU')

# Step 1: Load the Keras model
input_tensor = Input(shape=(224, 224, 3), name="input_layer")
model = ResNet101(weights="imagenet", input_tensor=input_tensor)

# Step 2: Convert the Keras model to OpenVINO IR format
print('started converting model')
ir_model = ov.convert_model(model)
print('finished converting model')

# Step 3: Compile the OpenVINO model
print('opening core')
core = Core()
print('compiling')
compiled_model = core.compile_model(
    model=ir_model,  # Use the converted model
    device_name="CPU",  # Specify the device (e.g., "CPU", "GPU")
)
print('finished compiling model')

# Step 4: Test the compiled model (Optional)
input_layer = compiled_model.input(0)
output_layer = compiled_model.output(0)

# Create a dummy input for testing
dummy_input = np.random.rand(1, 224, 224, 3).astype(np.float32)

# Perform inference
result = compiled_model([dummy_input])
print("Inference result:", result)

print("Model compilation and testing complete.")

When adding libtcmalloc.so.4 to LD_PRELOAD I get this:

started converting model
finished converting model
opening core
compiling
free(): invalid pointer

Process finished with exit code 134 (interrupted by signal 6: SIGABRT)

Same thing when adding libtbbmalloc_proxy.so to LD_PRELOAD.

[Jsy0220]

@rang-inspekto Hi, I think there may be a symbol conflict but not sure.
@ilya-lavrenov Hi, Could you reproduce it according to the sample offered by @rang-inspekto ? and how to solve it ?

[Jsy0220]

@rang-inspekto Hi, have you solved this problem ? I think I will use both tcmalloc and vino next too.

[rang-inspekto]

I'm also waiting for this to be fixed and currently workaround it by not using the additional malloc libs.

[Jsy0220]

@rang-inspekto okay, I have tried to ask them again.